Hear Me

Speech-to-text errors cluster in brittle terms: names, commands, flags, paths, identifiers, acronyms, versions, numbers, negation, and action words. Focus on a brittle term only when a wrong interpretation would change the action, target, scope, result, or safety. Accept ordinary wording as given.

Invocation

Match on and off without case sensitivity after trimming surrounding whitespace. Treat them as controls only when they are the entire argument of an explicit skill invocation. Reply to a control with only a confirmation of the new state.

• No argument or on: Enable dictation mode for the current conversation. Reply Hear Me is on for this conversation.
• off: Disable dictation mode. Reply Hear Me is off.
• Any other non-empty argument: Enable dictation mode and apply the protocol to that request. When this changes the state from off to on, state Hear Me is on for this conversation. after the Heard: readback and before other skill announcements or task work. Do not repeat the state for a later request while the mode is already on.

While the mode is on, apply the protocol to every later user message, including typed text and explicit skill invocations. Controls are the one exception: do not apply the protocol to an on or off control, and reply with only the confirmation. Only an explicit off control disables it. Start each new conversation with the mode off, and keep its state only in that conversation's history.

Protocol

1. Make a concise Heard: readback the first text of the first assistant message for each dictated request, before any skill announcement, tool call, or task work. Restate the action, target, scope, and all consequential constraints. Use one line when it preserves them; otherwise use a short block.
   • Format each correction or assumption as code. Include the original transcript when the change is not obvious: ruff check (from "rough check").
   • Mark an unresolved phrase and its credible candidates instead of choosing one: [unclear: feature/cash-sync or feature/cache-sync?].
   • Replace a secret or sensitive value with a role label such as [API token].
2. Resolve every consequential brittle term before work that depends on it. Use the conversation, repository state, and cheap read-only checks to test candidates. Treat a term as resolved only when one candidate has direct support and no credible alternative remains.
3. The readback is an error check, not a confirmation gate: proceed in the same turn when all dependent terms are resolved. When two or more credible interpretations remain, ask one focused question, complete every independent part of the request while that part is blocked, and mutate state only when the action and its target are both resolved.